Hi Martin, I agree with your theory. In the practical case of TProfile with low statistics in one bin and very small weights, do you have a better algorithm to propose? I mean an algorithm really working, and for example, not subject to rounding problems? Let me know. I will be happy to include it. Rene Brun On Wed, 17 Jul 2002, Martin Kestel wrote: > > Hi Rene, > > RB>The computation of errors in TProfile has evolved with time > RB>in case of bins with low statistics (1, 2 or 3 entries). > RB>In version 3.02/07, we introduced a new condition, in case > RB>the errors computed are very small (error/content <1e-6) > RB>We had complaints from several users making fits on such profiles > RB>and finding that too much weight was given to the points > RB>with low statistics. > RB>Any idea to improve the existing algorithm is welcome. > > as an old PAW user, I know that the error calculation within PAW has always > been the source of major trouble. > > Now, errors calculated for histograms have a meaning (usually). Making fits, I > need to be sure that the errors calculated for histograms are representative > of the data. If in a certain bin in profile histograms there are only few > entries, the error is going to be large (usually), independent of spread > option or error-on-the-mean option. > > Now, when the error is large, the point will get a low weight in a fit, that's > just how fits are set up to function. Therefore I can hardly understand how it > can happen that such data points get large weights. > > Introducing a fix or a fudge every time someone complains does not seem to be > a good policy; in the end we (as users) want to make serious physics with ROOT > and need to be able to rely on such basic things like error calculation. > > I do not see, why the spread option and the error-on-the-mean options shold > not suffice. They are meaningful and very (really!!!!) useful. > > > just my .02$ > > Martin >
This archive was generated by hypermail 2b29 : Sat Jan 04 2003 - 23:51:00 MET